Rate adaptation is a mechanism critical for maximizing the throughput of 802.11 systems. In this paper, inspired by the observation that the packet success rate under different data rates is not independent, we model the rate adaptation problem as a multi-armed bandit (MAB) problem with correlated arms. Two MAB-based rate adaptation algorithms are developed in which the correlation between data rates is exploited to accelerate the convergence of the algorithms. To verify the performance of our algorithms, we build up an indoor 802.11n test bed. The proposed rate adaptation algorithms are implemented and deployed on the test bed. Both simulation and test-bed experiments demonstrate the superiority of our algorithms in stationary and non-stationary radio environments.